PC-MAPPO: Phase-Conditioned Rollout Buffering for Multi-Robot Collaborative Motion Planning

Phase-dependent experience creates a practical tension between formation maintenance and obstacle-aware navigation in multi-robot motion planning. This paper presents PC-MAPPO, a phase-conditioned rollout framework for multi-robot formation and obstacle-aware navigation in the evaluated simulation setting. The method retains the Multi-Agent Proximal Policy Optimization (MAPPO) optimizer and reorganizes on-policy data through separate formation and navigation buffers, concatenated batch updates, and effective 0.5/0.5 transition weights whenever an update follows formation detection. Graph Attention Networks (GATs) encode a fixed scenario-specific interaction graph, while an Artificial Potential Field (APF)-inspired directional layer blends target and repulsive directions during obstacle-aware training and decentralized execution. The methodological contribution is an end-to-end phase-conditioned MAPPO formulation that carries an environment-derived phase label through rollout storage, phase-wise Generalized Advantage Estimation (GAE), advantage normalization, and loss aggregation, and couples this formulation to fixed-graph relational encoding and directional guidance. We evaluate the complete implementation in static fields with 0, 3, 5, and 7 obstacles, dynamic-obstacle scenes, and three- and four-agent narrow passages, using combined-implementation ablations. Additional multi-run ablations, coordination-pressure tests, topology comparisons, APF-mechanism comparisons, and static stress tests characterize the contribution, the conditions in which it is supported, and its operating boundaries.